Research And Implementation Of Intelligent Diagnostic Model Of Imbalanced Industrial Fault Under Variant Working Conditions

With the advent of the industrial 4.0 era,various information technology means to promote industrial innovation.Data is constantly accumulated in industrial systems,and data collected from controllers and sensors provides the basis for failure prediction and health management(Prognostics and Health Management,PHM)in industrial systems.PHM technology can use all kinds of data generated in industrial systems,through signal processing and data analysis and other computational means,to achieve the health of complex industrial systems to detect,predict and manage.Due to the accumulation of data,machine learning and deep learning techniques are applied to troubleshooting and predicting problems.Data-driven industrial troubleshooting methods are influenced by the characteristics of industrial data itself.Different operating conditions bring about differences in the distribution of industrial data,and the ratio of normal and fault data is unbalanced.The differences in data distribution and the unbalanced characteristics of fault data due to variable operating conditions can affect the performance of the troubleshooting model.Differences in data distribution under different operating conditions can lead to poor performance of well-trained troubleshooting models under certain operating conditions under new conditions.In summary,it is the differences in data distribution that challenge the generalization of the model.The unbalanced characteristics of industrial data make it difficult to learn the characteristics of a few types of fault samples,which makes it difficult for the model to classify fault samples correctly.In order to solve the above two challenges,this paper further studies theoptimization of industrial fault diagnosis model based on migration learning and unbalanced theory,the main results of which are as follows:1.In view of the differences in data distribution and the unbalanced characteristics of data brought about by variable conditions in industrial scenarios,a deep adversive transfer learning model(deep Imba-DA)based on cost-sentive loss is proposed to overcome these problems.In this method,cost-sensitive loss is used to solve the imbalance in fault diagnosis.Deep Imba-DA uses domain loss to minimize the marginal distribution difference between the source domain and the target domain and the maximum average difference(MMD)within the class to minimize conditional distribution differences.The performance of deep Imba-DA is evaluated and analyzed on the Case Western Reserve University(CWRU)dataset and the PADERBORN dataset.The results show that it is superior to other benchmark methods in bearing diagnosis.2.In view of the problem that the data under various operating conditions in the industrial scenario is not fully utilized in the migration learning model,this paper puts forward the fault diagnosis model(MSDA)based on multi-source domain migration learning.The model uses the corresponding source domain data under multiple operating conditions to learn,uses multiple sets of feature extractors,learns the characteristic information of multiple operating conditions,uses Wasserstein distance to measure the distribution discrepancy between the source domain and the target domain,and uses this distance to weight the classifiers corresponding to each group of feature extractors,and finally comprehensively outputs the results of fault diagnosis for the target conditions.In the course of training,the model makes full use of the bearing data collected under multiple operating conditions,which improves the generalization and classification accuracy of the model compared with the method of single-source migration.Finally,several indicators of the overall experiment are evaluated on the Case Western Reserve University(CWRU)bearing dataset.